Pippa: Okay, this is the clearest thing Anthropic’s written about Code in months.
Tyler: Right. Finally someone said it: model is the weights, effort is the loop.
Pippa: And it’s not just ‘think longer’—
Tyler: Mm-hm.
Pippa: it’s how many files it reads, how far it chases a refactor before it taps out. That’s the part I’ve been hand-waving for weeks.
Tyler: And the weights are literally frozen. So when it hallucinates an API that doesn’t exist—
Pippa: Yep.
Tyler: that’s the model, not the context. Context just steers within the weights.
Pippa: I love that split. It’s like they handed us the actual knobs.
Tyler: Though the ‘frozen’ framing bugs me a little…
Pippa: Of course it does.
Tyler: it’s technically accurate, but it sounds like they’re bragging about immutability.
Pippa: Tyler, it’s a blog post. They’re allowed to brag.
Tyler: Fair. And the defaults make sense—Opus 4.8 at default effort apparently matches 4.7’s results for similar token spend.
Pippa: That’s the kind of detail I’d bury in a parenthetical, but okay.
Tyler: It’s proof they’re not just cranking the effort knob to juice benchmarks.
Pippa: Speaking of which—wrong answer taxonomy: if it missed context it had, that’s model. If it didn’t even look, that’s effort.
Tyler: Exactly. And since all output is tokens—thinking, tool calls, the summary—
Pippa: Right.
Tyler: effort is literally buying more of that loop. No magic.
Pippa: This is the part that’s going to save me from my own impulse to tweak everything per-task.
Tyler: Impulse control is a feature.
Pippa: Says the guy who still hasn’t forgiven GraphQL.
Tyler: Unrelated. And still correct.
Pippa: Sure, sure. Anyway—the install command’s the same, right? curl -fsSL the link in the show notes
Tyler: Yep. And Desktop, VS Code, JetBrains, all the usual spots.
Pippa: Cool. So we’re done. This is episode six thirteen, and I’m weirdly relieved.